Routing easy jobs to cheaper models got Uber more than nine times the AI usage

Routing costs something before it saves anything. To send a job to a cheaper model you have to know the cheaper model does that job, which means an eval per task and a fallback for when it fails. Someone owns that eval the next time a vendor ships a new version. Uber's payoff, as the Fast Company column tells it, came only after it stopped treating every task the same ^[6].

The pressure that produced the change shows up in the token bill. A year's worth of tokens in four months is three times the rate the budget assumed ^[5][1].

The decision looks like a vendor choice: pick the best model and standardize on it. What the buyer in the column did was sort work by difficulty and pay frontier prices only for the hard part ^[6]. The column's wider claim is that companies will stop standardizing on one frontier model and start buying outcomes, at which point the model becomes an input to someone else's product ^[9].

On Anthropic, the column asserts more than it shows. It says many US customers are choosing cheaper models over the most advanced option, and that the company is finding it harder to persuade them to pay for its leading model, which the column calls Fable 5 ^[1][4]. It does not quantify how many moved or identify a single customer ^[12]. Both labs have also put cheaper models below the frontier while holding the price of their flagships ^[8].

For the person who has to write the routing policy on Monday, two axes do most of the sorting. One is whether a human reads the output before it leaves the team. The other is whether a wrong answer costs money you can put a figure on. Reviewed output with cheap errors goes to the cheapest model that passes your eval. Unreviewed output with expensive errors keeps the flagship until an eval says otherwise. Where a person checks the work but a mistake is costly, run the cheap model first and pay for the second look. Cheap errors with no reviewer get the cheap model and sampled audits.

I would set the default at the cheapest model that clears your eval, with a written reason on file for every workload that uses the flagship. The tradeoff is that you now maintain an eval suite and a router, and every model upgrade reopens both. The column's own prescription is aimed at the labs: sell portfolios at several price points, compete on the outcome, and let today's frontier model become tomorrow's cheap way in ^[10]. It expects the contest to move to the products and services wrapped around the models ^[13].